an invitation to toric degenerations - arxiv · an invitation to toric degenerations mark gross and...

arX

iv:0

808.

2749

v2 [

mat

h.A

G]

23

Jul 2

009

AN INVITATION TO TORIC DEGENERATIONS

MARK GROSS AND BERND SIEBERT

Contents

Introduction. 1

1. Purely toric constructions 2

2. Introducing singular affine structures 9

3. Examples without scattering 14

4. Scattering 23

5. Three-dimensional examples 34

References 41

Introduction.

In [GrSi2] we gave a canonical construction of degenerating families of varieties with

effective anticanonical bundle. The central fibre X of such a degeneration is a union of

toric varieties, glued pairwise torically along toric prime divisors. In particular, the notion

of toric strata makes sense on the central fiber. A somewhat complementary feature of

our degeneration is their toroidal nature near the 0-dimensional toric strata of X; near

these points the degeneration is locally analytically or in the etale topology given by a

monomial on an affine toric variety. Thus in this local model the central fiber is a reduced

toric divisor. A degeneration with these two properties is called a toric degeneration. The

name is probably not well-chosen as it suggests a global toric nature, which is not the case

as we will emphasize below. A good example to think of is a degeneration of a quartic

surface in P3 to the union of the coordinate hyperplanes. More generally, any Calabi-Yau

complete intersection in a toric variety has toric degenerations [Gr2]. Thus the notion of

toric degeneration is a very versatile one, conjecturally giving all deformation classes of

Calabi-Yau varieties with maximally unipotent boundary points.

This work was partially supported by NSF grants 0505325 and 0805328.

1

http://arxiv.org/abs/0808.2749v2

2 MARK GROSS AND BERND SIEBERT

Our construction has a number of remarkable features. It generalizes the construction

of a polarized toric variety from an integral polyhedron (momentum polyhedron) to non-

toric situations in a highly non-trivial, but canonical fashion. It works order by order,

each step being controlled by integral affine (“tropical”) geometry. This has the striking

consequence that any complex geometry feature of the degeneration that is determined on

a finite order deformation of the central fibre can, at least in principle, be read off tropically.

An important ingredient in this algorithm is the scattering construction, introduced by

Kontsevich and Soibelman in a rigid-analytic setup in dimension two [KoSo].

The purpose of these notes is to provide an extended introduction to [GrSi2]. The

emphasis is on highlighting some features of the construction by going through examples

explicitly. To avoid repeating ourselves, we will introduce most concepts in an ad hoc

fashion and refer to [GrSi2] for the general case and more technical definitions.

1. Purely toric constructions

1.1. Toric varieties from polyhedra. To start with let us recall the algebraic-geometric

construction of a toric variety from a convex integral polyhedron σ ⊆ MR, the intersection

of finitely many closed halfspaces. To keep track of functorial behaviour we work in

MR := M ⊗Z R for some free abelian group M ≃ Zn of rank n. If σ is bounded it is the

convex hull in MR of finitely many points in M ⊆ MR. In any case, for each face τ ⊆ σ

we have the cone generated by σ relative to τ :

Kτσ := R≥0 · (σ − τ) ={

m ∈ MR

∣

∣ ∃m0 ∈ τ,m1 ∈ σ, λ ∈ R≥0 : m = λ · (m1 − m0)}

.

These cones are finite rational polyhedral, that is, there exist u1, . . . , us ∈ M with

Kτσ = C(u1, . . . , us) := R≥0 · u1 + . . . + R≥0 · us.

Note also that Kτσ ∩ (−Kτσ) = Tτ = τ − τ , the tangent space to τ . So Kτσ is strictly

convex if and only if τ is a vertex.

The integral points of Kτσ define a subring C[Kτσ ∩ M ] of the Laurent polynomial

ring C[M ] ≃ C[z±11 , . . . , z±n

n ] by restricting the exponents to integral points of Kτσ. For

m ∈ Kτσ ∩ M we write zm for the corresponding monomial. The invertible elements of

this ring are precisely the monomials czm with m ∈ Λτ := Tτ ∩ M and c ∈ C \ {0}.

TORIC DEGENERATIONS 3

Example 1.1. Let σ ⊆ R2 be the triangle with vertices v1 = (0, 0), v2 = (1, 0), v3 = (0, a)

with a ∈ N \ {0}. Then

Kv1σ = C

(

(1, 0), (0, 1))

Kv2σ = C

(

(−1, a), (−1, 0))

Kv3σ = C

(

(0,−1), (1,−a))

For i = 1, 3 the monoid of integral points Kviσ ∩ Z2 ⊆ Z2 is freely generated by the

primitive generators (1, 0), (0, 1) and (0,−1), (1,−a) of the extremal rays. In other words,

N2 −→ Kv1σ ∩ Z2, (α, β) 7−→ α · (1, 0) + β · (0, 1) = (α, β)

and

N2 −→ Kv3σ ∩ Z2, (α, β) 7−→ α · (0,−1) + β · (1,−a) = (β,−α − βa)

are isomorphisms of additive monoids. This shows

C[Kviσ ∩ Z2] ≃ C[x, y], i = 1, 3,

as abstract rings. For i = 2 the integral generators (−1, a), (−1, 0) of the extremal rays of

Kv2σ generate a proper sublattice of Z2 of index a = det

(

−1 −1a 0

)

. Thus (−1, 0), (−1, a)

also do not suffice to generate Kv2σ ∩ Z2 as a monoid, for a > 1. It is not hard to show

that a minimal set of generators of Kv2σ rather consists of the a + 1 elements

(−1, 0), (−1, 1), . . . , (−1, a).

A good way to view C[Kv2σ ∩Z2] is as the ring of invariants of C[x, y] under the diagonal

action of Z/a by a-th roots of unity ζ ∈ C, ζa = 1:

x 7−→ ζ · x, y 7−→ ζ · y.

Under this identification z(−1,i) ∈ C[Kv2σ ∩ Z2] corresponds to the invariant monomial

xiya−i.

The remaining rings associated to higher dimensional faces of σ are

C[Kτσ] ≃

C[x, y±1], dim τ = 1

C[x±1, y±1], τ = σ.�

As the example indicates, rings of the form C[Kτσ ∩ M ] (toric rings) can be difficult

to describe in terms of generators and relations. To obtain examples that can be easily

written down in classical projective algebraic geometry, in this paper we therefore almost

exclusively restrict ourselves to polyhedra σ with Kvσ ∩ M ≃ Nn as a monoid, for any


vertex v ∈ σ. If m1, . . . ,ms are the generators of the extremal rays of Kvσ, a necessary

and sufficient condition for this to be true is s = n and det(m1, . . . ,ms) = 1.

Now given a convex integral polyhedron σ ⊆ MR, with dimσ = n for simplicity, and a

face τ ⊆ σ we obtain the affine toric variety

Uτ := Spec(

C[Kτσ ∩ M ])

.

Since C[Kτσ ∩ M ] ⊆ C[M ] any Uτ contains the algebraic torus Uσ = Spec(

C[M ])

≃ Gnm.

More generally, if τ ⊆ τ ′ then C[Kτσ ∩ M ] is canonically a subring of C[Kτ ′σ ∩ M ], and

hence we have an open embedding

Uτ ′ −→ Uτ .

These open embeddings are mutually compatible. Hence the Uτ glue to a scheme Xσ

of dimension dimσ. In other words, there are open embeddings Uτ → Xσ inducing the

morphisms Uτ ′ → Uτ for all τ ⊆ τ ′ ⊆ σ. The multiplication action on Uσ = Gnm extends

to Xσ. Hence Xσ is a toric variety. Note that according to this definition toric varieties

have a distinguished closed point, the unit of Gnm. Moreover, for faces τ ′ ⊆ τ ⊆ σ the ring

epimorphism

C[Kτ ′σ ∩ M ] −→ C[Kτ ′τ ∩ M ], zm 7−→

zm, m ∈ Kτ ′τ

0, otherwise,

induces a closed embedding ιτ : Xτ → Xσ with image disjoint from Gnm ⊆ Xσ unless

τ = σ. The images of the various Xτ are called toric strata of Xσ, the image of Uσ = Gnm

the big cell. If τ, τ ′ ⊆ σ are two faces it holds

ιτ (Xτ ) ∩ ιτ ′(Xτ ′) = ιτ∩τ ′(Xτ∩τ ′).

Here we make the convention X∅ := ∅. Hence the face lattice of σ readily records the

intersection pattern of the toric strata of Xσ . In particular, the facets (codimension one

faces) of σ are in one-to-one correspondence with the toric prime divisors, the irreducible

Weil divisors that are invariant under the torus action.

Example 1.2. For σ = conv{(0, 0), (1, 0), (0, a)} from Example 1.1 we claim that Xσ is the

weighted projective plane P(1, a, 1). Recall that P(1, a, 1) is the quotient of A3 \{

(0, 0, 0)}

by the action of Gm that on closed points is given by

λ · (x0, x1, x2) = (λx0, λax1, λx2), λ ∈ C∗.

In fact, on A3 \ V (x0) the ring of invariants of the action is C[x, y] with x = x1/xa0,

y = x2/x0 and similarly on A3 \ V (x2). On the other hand, on A3 \ V (x1) the ring of


invariants is generated by xi0x

a−i2 /x1, i = 1, . . . , a. This is the ring of invariants of the

diagonal Z/a-action on A2. Hence P(1, a, 1) has an affine open covering with spectra of

the rings C[Kviσ ∩ Z2] discussed in Example 1.1. The gluing morphisms between these

open sets are the same as given by toric geometry.

By construction the scheme Xσ depends only on the cones Kτσ, hence only on the

normal fan of σ with elements the dual cones (Kτσ)∨ ⊆ M∗R. More generally, toric

varieties are constructed from fans. In particular, integrality and boundedness of σ can

be weakened to rationality of the cones Kτσ. Those toric varieties coming from integral

polyhedra (bounded or not) are endowed with a toric ample line bundle. In fact, defining

the cone over σ

C(σ) := cl(

R≥0 · (σ × {1}))

⊆ MR × R,

the ring C[C(σ)∩ (M ×Z)] is graded by deg z(m,h) := h ∈ N. Taking the closure cl here is

important in the unbounded case. It adds the asymptotic cone lima→0 a · σ to MR × {0}.

It is then not hard to see that one has a canonical isomorphism

Xσ ≃ Proj(

C[C(σ) ∩ (M × Z)])

.

Although C[C(σ) ∩ (M × Z)] is not in general generated in degree 1, integrality of the

vertices of σ implies that the sheaf O(1) on the right-hand side is nevertheless locally free.

This yields the toric ample line bundle mentioned above.

1.2. Toric degenerations of toric varieties. Now let us see how certain unbounded

polyhedra naturally lead to toric degenerations with general fibre a toric variety. Let

σ ⊆ MR × R be an (n + 1)-dimensional convex integral polyhedron that is closed under

positive translations in the last coordinate:

σ = σ +(

0 × R≥0

)

.

Let q : MR × R → MR be the projection and

σ := q(σ).

Then the non-vertical part of ∂σ is the graph of a piecewise affine function

ϕ : σ → R

with rational slopes. The domains of affine linearity of ϕ define a decomposition P of σ

into convex polyhedra. In terms of this data, σ is the upper convex hull of the graph of ϕ:

σ ={

(m,h) ∈ MR × R∣

∣h ≥ ϕ(m)}

.


Thus σ is equivalent to a polyhedral decomposition P of the convex integral polyhedron

σ together with a function ϕ on σ that is piecewise affine and strictly convex with respect

to P and takes integral values at the vertices of P.

Now Xσ is an (n + 1)-dimensional toric variety that comes with a toric morphism

π : Xσ → A1.

In fact, each of the rings C[Kτ σ∩ (M ×Z)] is naturally a C[t]-algebra by letting t = z(0,1),

and the gluing morphisms are homomorphisms of C[t]-algebras. The preimage of the

closed point 0 ∈ A1 is set-theoretically the union of toric prime divisors of Xσ that map

non-dominantly to A1. It is reduced if and only if ϕ has integral slopes, that is, takes

integral values at all integral points, not just the vertices. To see this let v = (v, ϕ(v)) ∈ σ

be a vertex and ϕv(m) a piecewise linear function on MR which agrees with ϕ(v+m)−ϕ(v)

close to 0. In other words, the graph of ϕv(m) is the boundary of the tangent cone

Kvσ ={

(m,h) ∈ MR × R∣

∣h ≥ ϕv(m)}

of σ at v. A C-basis for C[Kvσ ∩ (M × Z)]/(t) is given by

z(m,h), ϕv(m) ≤ h < ϕv(m) + 1,

and z(m,h) is nilpotent modulo (t) if and only if ϕv(m) < h, that is, if ϕv(m) is not integral.

Assume now that ϕ(m) ∈ Z for all m ∈ σ ∩ M . Then C[Kvσ ∩ (M × Z)]/(t) has one

monomial generator z(m,ϕv(m)) for any m ∈ Kvσ ∩ M , and the relations are

z(m,ϕv(m)) · z(m′,ϕv(m′)) =

z(m+m′,ϕv(m+m′)), ∃τ ∈ P : v ∈ τ and m,m′ ∈ Kvτ

0, otherwise.

In other words, π−1(0) is the scheme-theoretic sum (fibred coproduct) of the n-dimensional

toric varieties

Xτ ≃ Xq(τ )

with τ ⊂ ∂σ projecting bijectively onto some τ ∈ P [n]. These are precisely the toric

prime divisors of Xσ mapping non-dominantly to A1. As in [GrSi2] P [k] denotes the set

of k-dimensional cells of the polyhedral complex P.

To understand general fibres π−1(t), t 6= 0, we localize at t. This has the effect of

removing the lower boundary of σ, that is, of going over to σ + (0 × R) = σ × R. Thus

π−1(A1 \ {0}) = Xσ×R = Xσ × (A1 \ {0}).

Thus each general fibre π−1(t), t 6= 0, is canonically isomorphic to Xσ. Note, however,

that these isomorphisms degenerate as t approaches 0.


Example 1.3. Here is a degeneration of P1 to two copies of P1 featuring an Al−1-

singularity in the total space. Let σ = [0, a+1], P [1] ={

[0, a], [a, a+1]}

, ϕ(0) = ϕ(a) = 0,

ϕ(a + 1) = l as in Figure 1.1. The slopes of ϕ are 0 and l on the two 1-cells.

The boundary of σ has two non-vertical components. Each gives one of the two ir-

reducible components of π−1(0). Their point of intersection is the 0-dimensional toric

stratum defined by the vertex v = (a, 0) of σ. The monoid Kvσ ∩ Z2 has generators

m1 = (−1, 0), m2 = (1, l), m3 = (0, 1) with relation m1 + m2 = l · m3. Hence,

C[Kvσ ∩ Z2] ≃ C[z1, z2, t]/(z1z2 − tl),

with zi = zmi for i = 1, 2 and t = z(0,1) defining π. This is a local model of a smoothing of

a nodal singularity with an Al−1-singularity in the total space. Thus the changes of slope

of ϕ at the non-maximal cells of P determine the singularities of the total space.

0 a a + 1

v

σ

(−l, 1)

Figure 1.1. The polyhedron σ (left) and the normal fan of σ (right).

Another way to understand the total space is from the normal fan of σ. It can be

obtained by subdividing the fan of A1 × P1 by the ray through (−l, 1). This corresponds

to a weighted blow up at one of the zero-dimensional toric strata, leading to another P1

over 0 ∈ A1 and the Al−1-singularity.

Note also that the length a of the interval is completely irrelevant to the complex

geometry; it only changes the polarization which we did not care about at this point. �

Example 1.4. For a two-dimensional example consider the convex hull

σ = conv{

(1, 0), (0, 1), (−1, 0), (0,−1)}

with the polyhedral decomposition into 4 standard 2-simplices shown in the figure below.

At (x, y) ∈ σ the function ϕ is given by 0, x, y and x + y, respectively, depending on the


maximal cell containing (x, y) as shown in the figure. Thus the slope of ϕ changes by 1

along each interior 1-cell of P.

The associated degeneration has as central fibre a union of 4 copies of P2, glued pairwise

along the coordinate lines as indicated by P. Thus the singular locus is a union of 4

projective lines, joined in one point Xv, where v = (0, 0, 0) ∈ σ. The monoid Kvσ is

generated by m1 = (1, 0, 1), m2 = (0, 1, 1), m3 = (−1, 0, 0), m4 = (0,−1, 0) fulfilling the

single relation m1 + m3 = m2 + m4. Thus

C[Kvσ ∩ Z3] ≃ C[z1, z2, z3, z4]/(z1z3 − z2z4),

with C[t]-algebra structure defined by t = z(0,0,1) = z1z3 = z2z4. This shows that the total

space Xσ has a singular point isomorphic to the origin in the affine cone over a smooth

quadric, while the central fibre is a product of two normal crossing singularities.

The general fibre is isomorphic to Xσ, as always, which here

is a toric Z/2-quotient of P1 × P1. In fact,(

1 1−1 1

)

maps the fan

of P1 × P1 to the normal fan of σ. Restricted to the big cell

G2m ⊂ P1 × P1, this map is given by

C[u±1, v±1] −→ C[x±1, y±1], u 7−→ xy, v 7−→ x−1y.

The subring of C[x±1, y±1] generated by x±1y±1 is the invariant

ring for the involution (x, y) → (−x,−y), and this involution ex-

tends to P1×P1. Note that Xσ has 4 isolated quotient singularities.

These correspond to the vertices of σ.

x0

y x + y

It is also instructive to write down this degeneration embedded projectively. As ex-

plained at the end of §1.1 we have to take the integral points of the cone C(σ) ⊆ R4 over

σ as generators of a graded C[t]-algebra, with the degree given by the projection to the

last coordinate and t = z(0,0,1,0). Now C[C(σ) ∩ Z4] is generated as a C[t]-algebra by the

monomials

X = z(1,0,1,1), Y = z(−1,0,0,1), Z = z(0,1,1,1), W = z(0,−1,0,1), U = z(0,0,0,1).

Note that these generators are in one-to-one correspondence with the integral points of σ.

The relations are

XY − tU2, ZW − tU2.

This exhibits Xσ as the intersection of two quadrics in P4A1 = A1 × P4. �


2. Introducing singular affine structures

From a birational classification point of view toric varieties are boring as they are all

rational. In particular, it is impossible to construct degenerations of non-rational varieties

directly by the method of §1.2. The idea in [GrSi2] is that one can get a much larger

and more interesting class of degenerations by gluing toric pieces in a non-toric fashion.

The central fibre is still represented by a cell complex P of integral polyhedra, but the

integral affine embedding of the cell complex into Rn exists only locally near each vertex.

In other words, the underlying topological space of P is an integral affine manifold B,

with singularities on a cell complex ∆ ⊆ B of real codimension 2 that is a retract of

|P [n−1]| \ |P [0]|. We then construct the total space X of the degeneration order by order,

by gluing torically constructed non-reduced varieties, “thickenings” of toric varieties so to

speak, in a non-toric fashion.

Thus our starting data are integral cell-complexes with compatible integral affine charts

near the vertices. We call these integral tropical manifolds ([GrSi2], Definition 1.2) because

they arise naturally as the bounded parts of the embedded tropical varieties associated to

the degeneration.

2.1. Degenerations of hypersurfaces. A hypersurface X ⊆ Pn+1 of degree d ≤ n +

2 can be degenerated to a union of d coordinate hyperplanes simply by deforming the

defining equation. For example, for n = 2 let f ∈ C[X0, . . . ,X3] be a general homogeneous

polynomial of degree d ≤ 4. Then

X0 . . . Xd−1 + tf = 0

defines a family π : Y → A1 with π−1(0) a union of d coordinate hyperplanes. This is

not a semistable family because Y is not smooth at the intersection of V (f) with the

singular locus of V (X0 · . . . · Xd−1). The latter consists of(

d2

)

projective lines V (Xi,Xj),

0 ≤ i < j ≤ d− 1. Since f is general the intersection of V (f) with any of these projective

lines consists of d reduced points with two nonzero coordinate entries each, that is, not

equal to the 4 special points [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]. Near any of these

points Y → A1 is locally analytically given by the projection of the three-dimensional

A1-singularity

V (xy − wt) ⊆ A4

to the t-coordinate. Note that for w 6= 0 this is a product of a semistable degeneration of

a curve with A1 \ {0}, but this fails at w = 0, which contains the singular point of Y .

While the local model of this degeneration is still toric, the singular points of Y are

general points of one-dimensional toric strata of π−1(0). Hence this is a very different


degeneration than the torically constructed ones in §1.2. Our first aim is to obtain this

local degeneration naturally from a tropical manifold. This was the starting point for

[GrSi2] in March 2004.

2.2. A singular affine manifold. There is a famous two-dimensional singular integral

affine manifold in the theory of integrable systems, called the focus-focus singularity [Wi],

which is the model for our singular affine manifolds B at general points of ∆, the codimen-

sion 2 singular locus of the affine structure. An (integral) affine structure on a topological

manifold is an atlas with transition functions in the (integral) affine linear group. Parallel

transport of tangent vectors is well-defined on such manifolds. In fact, an affine manifold

comes naturally with a flat, torsion-free, but usually non-metric connection. The focus-

focus singularity is (the germ at the origin P of) R2 with an integral affine structure away

from P such that parallel transport counterclockwise around P gives the transformation(

1 01 1

)

. This affine manifold B can be constructed by gluing R2 \ (R≥0 × {0}) with the

given affine structure to itself by the integral affine transformation

(x, y) 7−→ (x, x + y), for x ≥ 0, −x < y < 0,

see Chart I in Figure 2.1. Note that ∂y is an invariant tangent vector, and indeed the

projection (x, y) → x on R2 \ (R≥0 ×{0}) to the first coordinate descends to a continuous

map B → R that is integral affine away from P ∈ B. The preimage of the origin is a line

ℓ through P .

I II III

S

S

PPP

ℓ

Figure 2.1. Various charts I–III for the focus-focus singularity: On the

left, the shaded area is a fundamental domain for B. In the two other

figures, the dashed vertical lines indicate the parts of ℓ to be removed; the

horizontal curves are all straight affine lines.


Alternatively, B can be constructed from the two charts R2 \ ({0} × R≥0) and R2 \

({0} × R≤0), each covering B minus one half of ℓ, via

(

R \ {0})

× R −→(

R \ {0})

× R, (x, y) 7−→

(x, y), x < 0

(x, x + y), x > 0,

see Charts II and III in Figure 2.1. Note that we do not know how to continue any

non-vertical affine line across P .

2.3. The basic example. The focus-focus singularity admits a polyhedral decomposition

by decomposing along the invariant line ℓ. This decomposition has two maximal cells σ±,

with preimages R≥0×R and R≤0×R in any of the charts of Figure 2.1. We can also define

a strictly convex, integral affine function ϕ : B → R by letting ϕ = 0 on the left maximal

cell σ− and ϕ(x, y) = x on the right maximal cell σ+. Again, this is independent of the

chosen chart, and it takes integral values at any integral point of B.

Now let us try to treat this situation with a singularity in the same way as we did in

§1.2. We can still describe the central fibre as before:

X0 = Xσ−∐Xρ Xσ+

,

Xσ−= Spec

(

C[x,w±1])

, Xσ+= Spec

(

C[y,w±1])

, Xρ = Spec(

C[w±1])

,

where ρ = σ− ∩ σ+, x = z(−1,0), y = z(1,0), w = z(0,1). By writing this it is important to

fix one affine realization of σ±, say the Chart II in Figure 2.1. Intrinsically, the exponent

(−1, 0) defining x is an integral tangent vector on σ−, while the exponent (1, 0) defining

y is an integral tangent vector on σ+. For w we can take either maximal cell, because the

exponent (0, 1) just represents the global integral vector field ∂y defined above.

Continuing to work in Chart II the procedure of §1.2 yields the deformation V (xy−t) of

V (xy) ⊆ A2× (A1 \{0}). But in Chart III the tangent vector (1, 0) of σ+ changes to (1, 1).

Since z(1,1) = wy the same procedure applied in this chart therefore yields V (xy − wt)

for the deformation. Because w is invertible V (xy − t) and V (xy − wt) are isomorphic as

schemes over C[t], but not as deformation with the given embedding of X0. Of course, this

phenomenon is just due to the affine monodromy around P . As expected, the singularities

of the affine structure lead to inconsistencies in the naive approach using toric geometry.

The starting point to overcome this problem is to work modulo tk+1. In other words,

we want to construct a k-th order deformation Xk → Spec(

C[t]/(tk+1))

. The topological

spaces of X0 and Xk are the same, so we only have to deal with the structure sheaf.

We now use toric geometry merely to define the correct non-reduced versions (k-th order

thickenings) of the toric strata. We then glue two maximal-dimensional strata intersecting


in codimension one by a non-toric automorphism of the common codimension one stratum,

but the choice of this automorphism is different depending on which chart we use.

In the present example, the thickenings of the toric strata suggested by toric geometry

are given by the rings

Rkσ−

= Sk[x1, y1, w±1]/

(

x1y1 − t, yk+11

)

,

Rkσ+

= Sk[x2, y2, w±1]/

(

x2y2 − t, xk+12

)

,

Rkρ,σ−

= Sk[x1, y1, w±1]/

(

x1y1 − t, xk+11 , yk+1

1

)

,

Rkρ,σ+

= Sk[x2, y2, w±1]/

(

x2y2 − t, xk+12 , yk+1

2

)

,

with Sk = C[t]/(tk+1). Here we distinguish between x and y as monomials on σ− and on

σ+. Thus y1 = z(1,0,1) in either chart, while y2 = z(1,0,1) in Chart II and y2 = z(1,1,1) in

Chart III. These monomials only depend on the affine structure on σ± and hence have

an intrinsic meaning. For the thickening of the ρ-stratum, however, we obtain two rings,

depending on which maximal cell σ± the monomials live on. For the monomial w we don’t

need to make any choices because it corresponds to a globally defined vector field. In any

case, we have two natural Sk-algebra epimorphisms

q− : Rkσ−

−→ Rkρ,σ−

, q+ : Rkσ+

−→ Rkρ,σ+

,

exhibiting the thickened ρ-stratum as closed subscheme of the σ−- and σ+-stratum, re-

spectively.

Affine geometry suggests two isomorphisms hII, hIII : Rkρ,σ−

→ Rkρ,σ+

, depending on

which chart we use to go from σ− to σ+:

hII : x1 7−→ x2, y1 7−→ y2, w 7−→ w

hIII : x1 7−→ wx2, y1 7−→ w−1y2, w 7−→ w

Now comes the point: To remedy the inconsistency caused by this, let α ∈ C∗ and compose

hII with the automorphism

gII : x2 7−→ (α + w) · x2, y2 7−→ (α + w)−1y2, w 7−→ w

of the localization(

Rkρ,σ+

)

α+w. For Chart III we compose with

gIII : x2 7−→ (1 + αw−1) · x2, y2 7−→ (1 + αw−1)−1y2, w 7−→ w.

We then indeed obtain(

gII ◦ hII

)

(x1) = (α + w) · x2 = (1 + αw−1) · wx2 =(

gIII ◦ hIII

)

(x1),(

gII ◦ hII

)

(y1) = (α + w)−1 · y2 = (1 + αw−1)−1 · w−1y2 =(

gIII ◦ hIII

)

(y1).


The result of gluing SpecRkσ−

and SpecRkσ+

along the codimension one strata via this

isomorphism is given by the fibre product Rkσ−

×(Rkρ,σ+

)α+wRk

σ+. In this fibre product

the homomorphism Rkσ+

→ (Rkρ,σ+

)α+w is the composition of q+ with localization, while

Rkσ−

→ (Rkρ,σ+

)α+w composes q− and the localization homomorphism with gII ◦ hII =

gIII ◦hIII. It can be shown ([GrSi2], Lemma 2.34) that generators for this fibre product as

an Sk[w±1]-algebra are

x =(

x1, (α + w)x2

)

, y =(

(α + w)y1, y2

)

with the single relation (coming from x1y1 = t = x2y2)

(2.1) xy − (α + w)t = 0.

This is the k-th order neighbourhood of the surface degeneration discussed in §2.1, with

the A1-singularity at x = y = 0 and w = −α.

2.4. General treatment in codimension one. The discussion in §2.3 generalizes to

arbitrary dimension n as follows. Let ρ be a codimension one cell of P, separating the

maximal cells σ−, σ+. This gives three toric varieties Xσ±and Xρ with the respective big

cells Spec(

C[Λσ±])

≃ Gnm and Spec

(

C[Λρ])

≃ Gn−1m . For an integral polyhedron τ we use

the notation Λτ for the lattice of integral tangent vector fields on τ . The convex integral

piecewise affine function ϕ changes slope along ρ by some integer k > 0. We then obtain

the same rings and gluing morphisms as in §2.3, except that w has to be replaced by n−1

coordinates w1, . . . , wn−1 for Gn−1m ⊆ Xρ.

There are affine charts on the interiors of σ± and near each vertex of ρ. Each vertex

v ∈ ρ suggests an isomorphism hv : Rkρ,σ−

→ Rkρ,σ+

. For a different vertex v′ ∈ ρ we

have hv′(zm) = hv(z

m) for m ∈ Λρ because these tangent vectors are globally defined on

σ− ∪ σ+. But for general m ∈ Λσ−, parallel transport Πv′ to σ+ through v′ is related to

parallel transport Πv through v as follows:

Πv′(m) = Πv(m) + 〈m, dρ〉 · mρv′v.(2.2)

Here mρv′v is some element of Λρ, viewed canonically as a subset of Λσ+

, and dρ is the

primitive generator of(

Λv′/Λρ

)∗≃ Z that evaluates positively on tangent vectors pointing

from ρ into σ+. This monodromy formula implies

hv′(zm) =

(

zmρ

v′v

)〈m,dρ〉 · hv(zm).

Taking x = zm, y = z−m in the chart at v with 〈m, dρ〉 = 1 yields h′v(x) = zm

ρ

v′v · hv(x),

h′v(y) = z−m

ρ

v′v · hv(y). Thus to define the gluing invariantly we can compose hv with an


automorphism of Rkρ,σ+

of the form

gv : zm 7−→ f−〈dρ,m〉ρ,v · zm

for functions fρ,v ∈ C[Λρ], indexed by vertices v of ρ and fulfilling the change of vertex

formula

(2.3) fρ,v′ = zmρ

v′v · fρ,v.

In the example in §2.3 we can think of v as lying on the upper half of ℓ, v′ on the lower

half and fρ,v = 1 + αw−1, fρ,v′ = α + w. Then mρv′v = (0, 1), zm

ρ

v′v = w, and the change

of vertex formula reads α + w = w · (1 + αw−1). In the general case the result of gluing is

the hypersurface

xy − fρ,v(w1, . . . , wn−1)tk = 0.

3. Examples without scattering

In Section 2 we worked in an affine (in the algebraic geometry sense) neighbourhood of

a singular point of the total space of the degeneration. We now want to look at projective

examples in two dimensions, with one or several singular points. To this end we work with

bounded tropical manifolds and use the toric procedure from §1.2 for the treatment away

from the singularities.

3.1. Overview of the general procedure. To obtain degenerations of complete va-

rieties we need to start from a bounded tropical manifold (B,P), where B denotes the

underlying singular integral affine manifold and P is the decomposition into integral poly-

hedra, together with an (in general multi-valued) integral piecewise affine function ϕ. The

tropical manifold determines readily the prospective central fibre X0 =⋃

σ∈PmaxXσ from

the maximal cells, glued pairwise torically in codimension one. The most general approach

also allows one to compose the gluing in codimension one with a toric automorphism, but

as this is a rather straightforward generalization that only complicates the formulas we do

not include it here. In the notation of [GrSi2] this means taking se = 1 for any inclusion

e : ω → τ of cells ω, τ ∈ P.

To obtain local models near the singular points we need to choose, for each ρ ∈ P [n−1]

and vertex v ∈ ρ, a regular function fρ,v on an open set in Xρ ⊆ X0. The open set is the

affine neighbourhood

Uv := X0 \⋃

τ∈P, v 6∈τ

Xτ

of the corresponding point Xv ∈ X0. As explained in §2.4, consistency of the gluing

dictates the change of vertex formula (2.3) that relates fρ,v and fρ,v′ for vertices v, v′ ∈ ρ.


There is also a compatibility condition for the fρ,v for fixed v and all ρ containing a

codimension-2 cell that makes sure everything stays toric etale locally at general points of

the toric strata. We will explain this condition when we need it later on (see the discussion

after Example 4.1).

Next, the piecewise affine function ϕ determines local toric models for the degeneration,

defined by the toric procedure from §1.2. Explicitly, if v ∈ P is a vertex and ϕv is

a piecewise linear representative of ϕ introduced in §1.2 then the local model near the

zero-dimensional toric stratum Xv is

Spec(

C[P ])

−→ Spec(

C[t])

, P ={

(m,h) ∈ M × Z∣

∣ h ≥ ϕv(m)}

.

Here t = z(0,1) makes C[P ] into a C[t]-algebra. Following the discussion in Section 2 we

now work to some finite t-order k, that is modulo tk+1, and decompose according to toric

strata. This gives toric local models that depend on k, a toric affine open neighbourhood

Uω ⊆ X0 of Int(Xω), a toric stratum Xτ intersecting Uω and a maximal-dimensional

reference cell σ. In analogy with Uv for ω ∈ P we take

Uω := X0 \⋃

τ∈P, ω 6⊆τ

Xτ .

Thus we need σ ∈ Pmax and cells ω, τ ∈ P with ω ⊆ τ . The last condition is equivalent

to Uω ∩ Xτ 6= ∅. The corresponding ring, localized at all the gluing functions fρ,v with

ρ ⊇ τ , is denoted Rkω→τ,σ, see [GrSi2], §2.1 for details. Thus SpecRk

ω→τ,σ is a k-th order

non-reduced version (“thickening”) of (Xτ ∩ Uω) \⋃

ρ⊇τ V (fρ,v), v ∈ ω arbitrary. Note

that for v, v′ ∈ τ the gluing functions fρ,v and fρ,v′ differ by a monomial that is invertible

on Uω, so V (fρ,v′) ∩ Uω = V (fρ,v) ∩ Uω.

The relation ideal defining Rkω→τ,σ in (the localization at

∏

ρ⊇τ fρ,v of) C[P ] is generated

by all monomials z(m,h), (m,h) ∈ Λτ ⊕Z, that have τ -order at least k+1. Here one defines

the τ -order as the maximum, over all σ′ ∈ Pmax containing τ , of the order of vanishing

of z(m,h) on the big cell of Xσ′ , viewed canonically as a subset of Spec(

C[P ])

.

Now even in examples with few cells this procedure requires the gluing of many affine

schemes. We can, however, sometimes use the following shortcut. The integral points

of B provide a basis of sections of a natural very ample line bundle on the central fibre

X0, hence providing a closed embedding X0 → PN−1, N the number of integral points of

B. In sufficiently simple examples the gluing morphisms defining the deformation of X0

readily homogenize to describe the k-th order deformation of X0 as a subspace of PNOk

,

Ok = SpecSk, Sk = C[t]/(tk+1). The toric nature of this construction makes sure that


this procedure gives the right local models. This is how we compute most of the examples

in this paper.

3.2. One singular point I: Two triangles. For our first series of examples we fit

the basic example from §2.3 into a projective degeneration. This amounts to finding a

polyhedral decomposition of a bounded neighbourhood of the singular point P of the

focus-focus singularity in Figure 2.1. One of the easiest ways to do this is to take two

standard triangles (integrally affine isomorphic to conv{

(0, 0), (1, 0), (0, 1)}

), one on each

half-plane σ±, and with P at the center of the common edge. In particular, P should not

be integral, but rather half-integral. Thus let us first shift our Charts I–III by (0, 1/2) so

that the singular point P is at (0, 1/2). Now in Chart II take

σ1 = conv{

(0, 0), (−1, 0), (0, 1)}

, σ2 = conv{

(0, 0), (1, 0), (0, 1)}

as maximal cells for the polyhedral decomposition, see Figure 3.1. This is one of the

two choices making the boundary locally convex. In fact, we can always bring σ1 to

the suggested form by applying a transvection(

1 0a 1

)

on Chart I for some a ∈ Z. This

descends to an integral affine transformation of the focus-focus singularity. Then σ2 =

conv{

(0, 0), (1, b), (0, 1)}

for some b ∈ Z. Convexity at (0, 0) implies b ≥ 0, while convexity

at (0, 1) implies b ≤ 1 (check in Chart III). Local convexity of ∂B is indispensible to define

local toric models for the deformation. Both choices are abstractly isomorphic, b = 0

making the lower boundary a line, b = 1 the upper one.

IIIII

S

Sσ1 σ2

P

W

X Z Y

σ1

P

σ2

Z

W

X

Y

Figure 3.1. The two charts defining the tropical manifold B1.

There are 4 integral points on this tropical manifold B1, the vertices of P. We interpret

these as homogeneous coordinates in a P3 = Proj(

C[X,Y,Z,W ])

. The central fibre X0

is then given by the hypersurface XY = 0, because X and Y are the only vertices not

contained in the same cell. In fact, if P consists of standard simplices, X0 can be defined

by the Stanley-Reisner ideal [Ho],[St] of the simplicial complex given by P. This is clear

from the definitions.


Now the gluing computation in §2.3 readily homogenizes. In fact, viewing the compu-

tation in Chart II as being homogenized with respect to Z, the relation (2.1) becomes

X

Z

Y

Z− t(

α +W

Z

)

= 0.

Clearing denominators yields the family of quadric hypersurfaces

XY − tZ(αZ + W ) = 0.

The analogous computation in the other chart gives the same result. Hence we obtain

a projective degeneration with general fibre isomorphic to P1 × P1. The total space has

another A1-singularity of a toric nature at X = Y = Z = 0. This singularity is due to the

fact that ∂B1 is not straight at (0, 1).

It is interesting to compare this deformation to the torically constructed one for a

diagonally subdivided unit square with ϕ changing slope by 1 along the diagonal. As

W

Z

Y

≃

W

X

Y

Z X

Figure 3.2. Torically constructed degeneration of P1 × P1

shown in Figure 3.2 we again have 4 homogeneous generators X,Y,Z,W and the relation

XY − tZW = 0.

The change of coordinates W 7→ αZ + W shows that this family is isomorphic to the

previous one, the only difference being that now both A1-singularities of the total space

are at 0-dimensional toric strata of P3. Thus in a sense, by introducing the singular point in

the interior we have straightened the boundary of the momentum polytope at the vertex

labelled Z and moved one of the two A1-singularities in the total space to a non-toric

position. The fact that the introduction of a singular point leads to an isomorphic family

is a rather special phenomenon due to the large symmetry in this example.

3.3. One singular point II: Blow-up of P2. A similar tropical manifold, which we

denote B2, leads to a degeneration of the blow-up of P2 (Figure 3.3). Note that the

boundary of B2 is straight at both vertices labelled U and W . Again we take ϕ to change

slope by 1 along the only interior edge, with vertices labelled U and W . There are 5 integral

points, so we work in P4A1 = Proj

(

C[t][X,Y,Z,U,W ])

. Now σ2 is not a standard simplex.


( 1 01 1

)

S

S

σ2

σ1

σ1 σ2σ1

W σ2Y

Z

X U Y

W Z

X U

Figure 3.3. Three charts defining the tropical manifold B2.

Thus we have a relation that does not receive any corrections from monodromy, namely

WY − UZ = 0.

The gluing computations at the vertices labelled U and W homogenize to the two equations

XY − t(αU + W )U = 0,

XZ − t(αU + W )W = 0.

For UW 6= 0 these are related by the substitution Z = WY/U . The central fibre X0 has

two irreducible components P2 and P1 × P1, glued along a P1. The total space has only

one singular point, the A1-singularity at [0, 0, 0, 1,−α] ∈ (X0)sing = P1. The general fibre

of this degeneration π : X → A1 contains the (−1)-curve E = V (X,αU + W,αY + Z)

whose contraction yields a P2. This is a well-known example: The 3 relations form the

2 × 2-minors of the matrix(

U W X

Y Z t(αU + W )

)

,

which for t = 1 describe the cubic scroll, see e.g. [Ha], Example 7.24.

Again, this example has a toric analogue (Figure 3.4). The relations are

WY − UZ = 0, XY − tU2 = 0, XZ − tUW = 0.

In fact, the substitution U 7→ αU + W , Y 7→ αY + Z yields the non-toric ideal above.

3.4. Propagation. Once we start changing the gluing of components somewhere we are

forced to change at other places as well to keep consistency. Thus in a sense the gluing

functions propagate.

Example 3.1. Consider the non-compact example of a tropical manifold with 4 maximal

cells σ1, . . . , σ4 sown in Figure 3.5. If we take the piecewise affine function ϕ with ϕ(1, 0) =


Y

Z

WX

U≃

UX Y

ZW

Figure 3.4. Torically constructed degeneration of P2 blown up once

σ2

σ3

σ1

σ4

v

ρ3

ρ4

ρ2

ρ1

xy

w

z

( 1 10 1

)

Figure 3.5. A tropical manifold forcing propagation of the gluing func-

tion. The figure on the right shows the tangent vectors belonging to the

generators of the toric model.

1, ϕ(0, 1) = 1, ϕ(−1, 0) = 0, ϕ(0,−1) = 0, the toric local model at v is

C[t] −→ C[x, y, z, w]/(xy − zw), t 7−→ xy = zw.

In terms of integral tangent vectors at v the generators are

x = z(1,0,1), y = z(−1,0,0), z = z(0,1,1), w = z(0,−1,0).

Now let us try to glue together the k-th order neighbourhood of X0 following the general

procedure of §2.4 with gluing function fρ3,v = 1 + y and all other fρi,v = 1. We have

4 rings Rkv→σi,σi

defining the thickenings of the irreducible components Xσi, 4 · 2 = 8 rings

for the one-dimensional toric strata Rkv→ρi,σi

, Rkv→ρi,σi−1

(i taken modulo 4), and 4 rings,

identified mutually via an affine chart at v,

Rkv→v,σi

= C[x, y, z, w, t]/(

(xy − zw, xy − t) + Ik

)

for the thickening of the zero-dimensional toric stratum Xv. Here Ik = (x, z)k+1 +

(x,w)k+1 + (y, z)k+1 + (y,w)k+1 is the ideal generated by monomials that are divisible

by tk+1 at the generic point of Xσifor some i. For the toric local model we want to take


the inverse limit of all these rings with respect to the homomorphisms that we have intro-

duced between them. These are of two kinds. First, for τ ′ ⊆ τ we have Rkv→τ,σi

→ Rkv→τ ′,σi

defining the inclusion of toric strata with reference cell σi. Second, there are the change

of reference cell isomorphisms Rkv→τ,σi

→ Rkv→τ,σi±1

. This requires compatibility of the

compositions. In the present case this comes down to checking the following. Let θkv be

the composition

Rkv→v,σ1

−→ Rkv→v,σ2

−→ Rkv→v,σ3

−→ Rkv→v,σ4

−→ Rkv→v,σ1

of changing the reference cell σi counterclockwise around the origin by crossing the ρi.

The compatibility condition is θkv = id. But crossing ρ1, ρ2 and ρ4 yields the identity.

Thus θkv equals the change of reference cell isomorphism from crossing ρ3, which is

x 7→ x, y 7→ y, z 7→ (1 + y)z, w 7→ (1 + y)−1w.

Recall that we pick up a negative power of fρ3,v = 1 + y if the tangent vector points from

ρ3 into σ3, while monomials with exponents tangent to ρ3 are left invariant. Note also

that because y is nilpotent now, 1 + y is invertible, so this is a well-defined automorphism

of Rkv→v,σ1

. In any case, we see θkv 6= id as soon as k ≥ 2.

To fix this it is clear how to proceed: Just take fρ1,v = 1 + y as well. Then by the

sign conventions, crossing ρ1 gives the inverse of what we had for ρ3. Then θkv = id as

automorphism of Rkv→v,σi

for any k (and any i). �

Remark 3.2. There is an important observation to be made here. On Xρ3the monomial y

is nonzero, that is, of order zero with respect to t. But once we move to a general point of

Xρ1we have y = x−1t because of the relation xy = t. Thus the t-order increases. And this

is not an accident, but happens whenever we follow a monomial z(m,h), (m,h) ∈ Λv × Z

in direction −m ([GrSi2], Proposition 2.6). More precisely, if ϕ differs by aρ ∈ Λ⊥ρ along a

codimension one cell ρ then a monomial with tangent vector m passing through ρ changes

t-order by −〈m,aρ〉.

This implies that even in examples where the invariant line ℓ emanating from a singular

point is unbounded, the gluing functions always converge t-adically. Said differently, for

the construction of a k-th order deformation of X0, we need to propagate the contribution

of each singular point only through finitely many cells. �

Of course, there is no reason for the propagation to stay in the 1-skeleton of P. The

gluing functions then separate from the (n− 1)-skeleton of P and move into the interiors

of maximal cells.


Example 3.3. Let us modify Example 3.1 as shown in Figure 3.6 with ϕ(−1, 0) = 0,

ϕ(0,−1) = 0, ϕ(1, 1) = 1. The gluing functions are fρ2,v = 1 + x and all other fρi,v = 1.

The toric local model at v is now the normal crossing degeneration

σ2

σ3

ρ2

ρ3

x

y

z

ρ1

v

p

σ1

( 1 10 1

)

Figure 3.6. Propagation of the gluing function into a maximal cell. The

figure on the right shows the tangent vectors belonging to the generators

of the toric model.

C[t] −→ C[x, y, z], t 7−→ xyz

with x = z(−1,0,0), y = z(0,−1,0), z = z(1,1,1). Following the gluing isomorphisms around v

leads to the automorphism

θkv : x 7→ x, y 7→ (1 + x)−1y, z 7→ (1 + x)z

of the ring

Rkv→v,σi

= C[x, y, z, t]/(

xyz − t, xk+1, yk+1, zk+1)

.

For k > 0 this is again not the identity, and we can’t fix this by changing any of the gluing

functions. But since 1 + x is invertible on the whole thickening Spec(

C[x, y, z, t]/(xyz −

t, xk+1))

of Xσ1, the inverse of θk

v can be viewed as an automorphism of Xσ1that propagates

along the ray p = R≥0 · (1, 0) emanating from the codimension two cell v. We then change

the gluing procedure by taking one copy of the thickening of Xσ1for each connected

component of σ1 \ p. Passing through p means applying the attached automorphism. The

thickenings of the codimension one strata Xρ1and Xρ3

are viewed as closed subschemes of

the copies labelled by the connected component of σ1\p containing ρ1 and ρ3, respectively.

Inserting p makes the gluing consistent to all orders. The result is isomorphic to

Spec(

C[x, y, z, t]/(xyz − (1 + x)t))

. �

Apart from the gluing functions we have now introduced another object, automorphisms

propagating along rays. The higher-dimensional generalization of rays are walls ([GrSi2],

Definition 2.20): They are one-codimensional polyhedral subsets of some maximal cell


σ, emanating from a two-codimensional polyhedral subset q into the interior of σ, and

extending all the way through σ along some integral tangent vector −mp, that is, p =

σ ∩ (q − R≥0 · mp). The attached automorphism is of the form

z(m,h) 7−→(

1 + cpz(mp ,hp)

)〈m,np〉 · z(m,h),

where hp > 0 is the t-order of the attached monomial and cp ∈ C. We will discuss how to

obtain these walls systematically in the context of scattering in Section 4.

3.5. Several singular points: The non-interacting case. For several singular points

P1, . . . , Pl ∈ B the monodromy invariant direction determines an affine line ℓµ emanating

from each Pµ. As long as these coincide or do not intersect, the construction presented so

far works.

Example 3.4. Here is an example with two singular points having parallel invariant lines

ℓ1, ℓ2. We will see that, depending on a choice of parameter, it leads to a degeneration

S S

S Sσ1

X U

R S

YV

σ3

σ1 σ3

ℓ1

σ2

ℓ2

ℓ1 ℓ2

σ2

X R S Y

VU

Figure 3.7. Two charts for a tropical manifold with two parallel invariant

lines ℓ1, ℓ2.

either of the Hirzebruch surface F2 = P(

OP1(−2) ⊕OP1

)

or of P1 × P1.

In this case we have two codimension one strata along which we glue. Taking ϕ to

change slope by 1 along each interior edge again, the homogenization of the gluing along

codimension one cells gives the polynomials

XV − t(U + R)U, XS − t(R + U)R, UY − t(V + λS)V, RY − t(λS + V )S.

Here λ ∈ C∗ is a parameter that we can not get rid of by automorphism. Note that λ

determines the relative position of the zero loci of the gluing functions when compared in

Xσ2= P1×P1. There is one more toric relation RV −SU = 0 from σ2. The corresponding

subscheme of P5A1 = ProjC[t]

(

C[t][X,Y,R, S,U, V ])

has one extra component V (R,S,U, V ).

In fact, saturating with respect to any of R,S,U, V gives another relation XY − t2(U +

R)(V + λS) = 0. This relation can also be deduced directly from following the line

connecting the vertices X and Y in the chart shown in Figure 3.7 on the right.


These six polynomials are the 2 × 2 minors of the matrix(

X t(V + λS) U R

t(U + R) Y V S

)

For fixed t 6= 0, this is a scroll in P5 whose ruling is given by the lines whose equations are

linear combinations of the two rows of the above matrix. The image of this scroll under

the rational map κ : Xt → P3 given by [x, y, u, v, r, s] 7→ [x0, x1, x2, x3] = [u + r, y, v, s] has

equation X0X1− t(X2 +X3)(X2 +λX3) = 0 in P3. For λ 6= 1 this is a smooth quadric and

κ induces an isomorphism Xt ≃ P1 × P1. For λ = 1 the quadric X0X1 − t(X2 + X3)2 = 0

has an A1-singularity, κ contracts the (−2)-curve given by Y = V + S = U + R = X = 0,

and Xt ≃ F2.

Example 3.5. This example features two singularities with the same invariant line. The

σ1

σ2

σ3 σ4 σ3

σ1

σ2

σ4

Q S

Figure 3.8. A tropical manifold for a degeneration of P2 blown up twice.

The invariant lines emanating from the two singular points agree.

general fibre of the corresponding degeneration is a P2 blown up in two different points.

We don’t bother to write down the homogeneous equations.

This example is easy to generalize to any number of blown up points, which then gives

our first genuinely non-toric examples.

Many more examples can be obtained from [Sy], which contains a toolkit for the con-

struction of two-dimensional affine manifolds with focus-focus singularities. The sym-

plectic 4-manifold constructed in this reference from such a singular affine manifold is a

symplectic model for the general fibre of our degeneration.

4. Scattering

If the invariant lines emanating from singular points of a two-dimensional affine struc-

ture intersect, the gluing construction becomes inconsistent again.


4.1. First example of scattering. We study the modification of Example 3.1 shown in

Figure 4.1, with gluing functions

fρ1,v = 1 + y = 1 + x−1t, fρ2,v = 1 + w = 1 + z−1t, fρ3,v = 1 + y, fρ4,v = 1 + w.

Following the gluing isomorphisms of Rkv→v,σi

by going counterclockwise around v, with

σ2

σ3

σ1

σ4

v

ρ3

ρ4

ρ2

ρ1

xy

w

z

( 1 10 1

)

(

1 0−1 1

)

Figure 4.1. A tropical manifold leading to inconsistent gluing. The figure

on the right shows the tangent vectors belonging to the generators of the

toric model.

starting cell σ1 say, yields

x 7−→ (1 + w)x

7−→(

1 + (1 + y)−1w)

x

7−→(

1 +(

1 + (1 + w)y)−1

w)

(1 + w)−1x

7−→(

1 +(

1 + (1 + (1 + y)w)y)−1

(1 + y)w)

(

1 + (1 + y)w)−1

x

=(

1 +(

(1 + y) + (1 + y)wy)−1

(1 + y)w)

(

1 + (1 + y)w)−1

x

=(

1 + (1 + wy)−1w)

(1 + w + wy)−1x

=(

(1 + wy) + w)

(1 + wy)−1(1 + w + wy)−1x = (1 + wy)−1x,

and similarly,

y 7−→ (1 + wy)y, z 7−→ (1 + wy)z, w 7−→ (1 + wy)−1w.

Thus we can again achieve θkv = id for all k by inserting the ray p = R≥0 · (1, 1) with

attached function 1 + wy = 1 + t2x−1z−1!

4.2. A projective example. Of course, this example can also be made projective, say

as in Figure 4.2. The ideal defining the degeneration is generated by the two polynomials

XY − t(U + W )U, ZW − t(U + Y )U.


Up

Y X

W

Z

Figure 4.2. Chart for the tropical manifold B3. The dashed lines denote cuts.

These relations can be most simply derived from gluing k-th order standard thickenings

of Xσ3, Xσ4

and of Xσ2, Xσ3

, respectively, and let k tend to infinity. This procedure does

not seem to depend on p. In fact, in this example, the insertion of p merely assures that

the same relations are obtained with any other pairs of neighbouring maximal cells.

As an abstract deformation this is isomorphic to the toric one with ideal (XY −

tU2, ZW − tY 2) presented in Example 1.4.

4.3. Systematic procedure for the insertion of walls. The long computation for

the insertion of the ray p in Example 4.1 looks like quite an accident. But there is a

systematic procedure to insert rays (or walls in higher dimension) to achieve consistency

to any finite order. The number of rays to be inserted becomes arbitrarily large with

increasing k, however, essentially with the only exception the example discussed in §4.1.

This procedure originates from the two-dimensional situation of [KoSo]. Of course, due to

the non-discrete nature of the discriminant locus the higher dimensional situation is much

more involved.

Systematically we proceed as follows. By induction one arrives at a collection of walls

and modifications of the gluing functions, a structure (in its architectural meaning) S as

we call it. A structure consists of walls and so-called slabs. A slab b is a polyhedral subset

of some ρ ∈ P [n−1] together with a collection fb,v of higher order corrections of the gluing

functions fρ,v, one for each vertex v ∈ ρ that is contained in a connected component of

ρ \ ∆ intersecting b. These fulfill the same change of vertex formula (2.3) as the fρ,v, but

interpreted in the rings Rkρ→ρ,σ for any σ ∈ Pmax containing ρ. We need slabs because

a wall hitting a codimension one cell ρ may lead to different corrections on both sides of

the wall. Technically we refine P to a polyhedral decomposition PS in such a way that

the walls lie in the (n − 1)-skeleton of PS . Then the underlying subsets of the slabs are

exactly the (n − 1)-cells of PS contained in the (n − 1)-skeleton of P.


For consistency of the gluing construction it suffices to follow loops around the codi-

mension two cells of PS , the joints of the structure (or rather of PS ). This is shown in

[GrSi2], Lemma 2.30; the precise definition of consistency is in Definition 2.28. The fact

that the gluing construction for a structure consistent to order k gives a k-th order toric

degeneration occupies §2.6 in [GrSi2].

Thus consistency is really a codimension two feature. For the scattering computation

at a joint j we can therefore work over the ring S = C[t℄[Λj] of Laurent polynomials with

coefficients in C[t℄ and exponents in Λj, the lattice of integral tangent vectors along j. In

fact, if θj is the composition of the gluing morphisms following a small loop around such

a joint, then θj is the identity on any monomial z(m,h) with m tangent to j.

By making the discriminant locus ∆ sufficiently non-rational or by deforming it a little,

we can assume the intersection of ∆ with j to be of codimension one in j. Let p ∈ j \ ∆

be a general point and denote by σj ∈ P the smallest cell containing j. Note that σj has

dimension n − 2, n − 1 or n. The computations take place in an affine chart at a vertex

v ∈ σj in the same connected component of σj \∆ as p. In an affine chart at v decompose

the lattice of integral tangent vectors by choosing a complement Λ ≃ Z2 to Λj ⊆ Λv

Λv = Λj ⊕ Λ.

While the procedure essentially is always the same, the setup for the scattering differs

somewhat depending on dimσj ∈ {n − 2, n − 1, n}. For being explicit let us assume

dim σj = n − 2, which is is the most difficult case; it is also the situation for the initial

scattering of codimension one cells of P intersecting in a codimension two cell j = σj of

P. At a later stage (k > 0) the inclusion j ⊆ σj may be strict and there may be additional

walls running into j from scatterings at other joints.

The scattering procedure is a computation in the rings Rkσj→σj ,σ

using induction on k.

However, these rings are not practical for computations with a computer algebra system,

and they would require the introduction of log structures in Example 4.1. Log structures

are necessary to make things work globally in the end, but on the level of this survey it

does not seem appropriate to get into these kinds of technicalities. We therefore present

the scattering computation in a traditional polynomial ring.

For any order k′ ≥ 0 the composition of gluing morphisms around j, starting at a

reference cell σ ∈ Pmax, defines an automorphism θk′

jof the ring Rk′

σj→σj ,σ. This ring is

a finite Sf -algebra, where f =∏

ρ⊇σjfρ,v. Let ϕ be the restriction to {0} × Λ ≃ Z2 of a

non-negative representative of ϕ. By changing the representative by an affine function we

can assume

ϕ(0, 0) = ϕ(−1, 0) = 0, 0 ≤ ϕ(0, 1) ≤ ϕ(1, 1),


which by convexity and integrality of ϕ implies

ϕ(1, 1) ≥ 1.

Letting x := z(−1,0,0), y := z(0,−1,0), z := z(1,1,1) this exhibits Rk′

σj→σj ,σas a quotient of an

Sf -subalgebra of

Rk := Sf [x, y, z]/(

(xyz − t) + (x, y, z)k+1)

,(4.1)

for some k ≥ k′. The smallest possible k is the minimum of the degrees a + b + c for

xaybzc having σj-order (introduced in §3.1) larger than k′. Conversely, any computation

in Rk can be obtained from a computation in Rk′

σj→σj ,σfor k′ ≫ 0. Thus the result of the

scattering procedure is the same with both sorts of rings.

Initially the only scattering happens at codimension two cells τ ∈ P. The starting

data are the gluing functions fρ,v for all codimension one cells ρ ⊇ τ where v ∈ τ is

some vertex selecting an affine chart for the computation. Consistency for k = 1 already

requires a compatibility condition for the restrictions fρ,v|Xτ that we can see by studying

the following local example.

Example 4.1. To consider a local scattering situation for k = 1 we take j = τ = Rn−2 ×

{0} ⊆ Rn as an affine subspace and ρi = τ × R≥0mi, i = 1, . . . , l, affine half-hyperplanes

emanating from τ in directions mi ∈ Z2 = Λ in the normal lattice Zn/Λτ = Λ. (Strictly

speaking this is not a legal situation because τ does not have a vertex, but this is irrelevant

for the present discussion.) We assume the ρi labelled cyclically around τ , that is, ρi

separates two maximal cell σi−1, σi containing τ , for any i ∈ Z/lZ. Let ni = (ai, bi) be

a primitive normal vector to Λρithat evaluates non-negatively on σi. Then passing from

σi−1 to σi gives the automorphism

θ1i : zm 7−→ f−〈ni,m〉

ρi· zm

of the ring R1 = Sf [x, y, z]/(x, y, z)2 from (4.1). Because in this local example there is no

vertex we drop the reference vertex from the notation for the gluing functions fρi. The

loop around the joint τ gives the automorphism θ1τ = θ1

k ◦ . . . θ11. Because we work modulo

(x, y, z)2 the effect of applying θik to a monomial is multiplication with a power of the

restriction fρi|Xτ ∈ Sf . We obtain

θ1τ (x) =

(

∏

i

f−〈(ai,bi),(−1,0)〉ρi

)

· x =(

∏

i

(

fρi|Xτ

)ai

)

· x,

and analogously,

θ1τ (y) =

(

∏

i

(

fρi|Xτ

)bi

)

· y, θ1τ (z) =

(

∏

i

(

fρi|Xτ

)−ai−bi

)

· z.


Thus consistency for k = 1 requires the following multiplicative condition for the slab

functions

∏

i

(

fρi|Xτ

)bi = 1 and∏

i

(

fρi|Xτ

)ai = 1.(4.2)

�

Example 4.1 shows that apart from the change of vertex formula (2.3), slab functions

need to fulfill a number of algebraic equations to achieve consistency for k = 1. More

precisely, each τ ∈ P [n−2] yields two multiplicative equations as in (4.2) for the restrictions

fρ,v|Xτ of the slab functions for those ρ ∈ P [n−1] that contain τ , see Equation (1.10) in

[GrSi2] for the general form. Note that by the change of vertex formula the relations

obtained from different vertices v, v′ ∈ τ are equivalent.

For those readers familiar with [GrSi1], what this shows is that specifying the gluing

functions fρ,v for k = 1 is essentially the same as specifying a log structure on X0 (cf.

[GrSi1], Theorem 3.22). Thus the log structures of [GrSi1] on X0 can be viewed as “initial

conditions” for our construction.

To explain the next steps (k > 1) let us look at the following typical scattering situation

in dimension 3. Let the reference vertex be v = 0 ∈ R3 and let the codimension two cell

τ be contained in the coordinate axis R · (0, 0, 1). We have three codimension one cells

ρ1, ρ2, ρ3 ⊇ τ with Kτρi/Tτ spanned by (−1, 0, 0), (−1,−3, 0) and (2, 3, 0), respectively.

For each i there is a unique slab bi ⊆ ρi containing x. The piecewise affine function ϕ is

uniquely defined by ϕ|τ = 0 and

ϕ(−1, 0, 0) = ϕ(−1,−3, 0) = 0, ϕ(2, 3, 0) = 3.

Following the maximal cells σ1, σ2, σ3 containing τ counterclockwise, with σ1 containing ρ1

and ρ2, we have ϕ(a, b) = 0, 3a− b and b for (a, b) ∈ σ1, σ2, σ3, respectively. In particular,

ϕ(2, 3, 0) = 3 is the smallest possible value making ϕ integral and with ϕ|σ1= 0. The

exponents of our monomials take values in

P :={

(m,h) ∈ Λ × Z∣

∣h ≥ ϕ(m)}

.

As slab functions take

fb1,v = 1 + z(0,0,1,0) + 2z(−1,0,0,1),

fb2,v = 1 + z(0,0,1,0) − z(−1,−3,1,0),

fb3,v = 1 + z(0,0,1,0) + 5z(2,3,0,3),


where the exponents lie in Λ×Z, Λ = Λv. Thus in z(a,b,c,h) the first two coordinates (a, b)

determine the direction in Λ/Λτ , the third coordinate c is for the toric parameter on Xτ

and h − ϕ(a, b) determines the divisibility by t.

We also assume there is one wall p containing τ coming from some earlier step, say with

Kτp/Tτ spanned by (−1, 1, 0) and with associated function

fp = 1 + 7z(−1,1,0,2).

Our scattering problem can be visualized in the two-dimensional normal space to τ as in

Figure 4.3 on the left.

b1

b2

b3

p

4c 0c 10c

4a 10a

4b

10b

6a

0a

8a

0b

8b

8c

p

Figure 4.3. A scattering procedure: Starting data (left) and result at

order 10. The labels are explained in the text.

For the explicit scattering computation we now go over to the Sf -algebra Rk from (4.1).

Here Sf = C[t℄[w,w−1]1+w with w = z(0,0,1,0), and

x = z(−1,0,0,0), y = z(0,−1,0,0), z = z(1,1,0,1).

Note that xyz = t defines the t-algebra structure, and a monomial xaybzcwd represents

an element of P if and only if

ϕ(c − a, c − b, d) = max{

0, c − b, 3(c − a) − (c − b)}

≥ c.

With these notations our input data is

fb1,v = 1 + w + 2tx, fb2,v = 1 + w − wxy3, fb3,v = 1 + w + 5xz3, fp = 1 + 7tx2z.


It is now easy to compute the composition θkv by a computer algebra system. Seeing the

various terms come up at higher order and getting rid of them by insertion of walls is

very instructive. We therefore encourage the reader to do that and verify the following

computations. Note that as we have seen in §4.1, the naive insertion procedure leads

to huge expressions quickly, so a computation in Q[x, y, z, w] would be too slow. It is

therefore important to expand in a Taylor series to order k in each of x, y and z after each

application of an automorphism associated to a slab or wall. Note also that to compute

θkv as an automorphism of Rk

τ→τ,σ1we have to observe the difference between the τ -order

of a monomial xaybzc and its degree a + b + c. For example, both x2y and x5y5z5 = t5

have τ -order 5, but x2y is already visible modulo (x, y, z)4 while x5y5z5 requires working

modulo (x, y, z)16.

The computation in Rk gives consistency up to k = 3. At degree 4 one has to insert

three walls, denoted 4a, 4b, 4c in Figure 4.3 on the right. The associated functions are

1 + 5(1 + w)2xz3, 1 + 2(1 + w)2x2yz, 1 − (1 + w)2wxy3.

This is just the part fbi− (1 + w) of the slab functions that do not cancel, with an

appropriate power of 1+w that we will explain shortly. Similarly, at degree 6 the incoming

wall p starts contributing, forcing its continuation by a wall in direction (1,−1) with

function 1 + 7(1 + w)5tx2z (6a in Figure 4.3 on the right). Then everything goes well up

to degree 8, where we find interactions of two of such terms, namely some multiples of

x3yz4 = tx2z3, x3y4z = tx2y3, x2y3z3 = t2yz. These monomials have directions (1, 3),

(−2,−3), (1, 0), so have to scatter away in the directions (−1,−3), (2, 3), (−1, 0) occupied

by the slabs. Since we do not want to allow walls to lie above slabs the only possibility is

to change the slab functions. Here is the change that makes things consistent to degree 9:

f ′b1,v = 1 + w + 2tx − 15(1 + w)2t2wyz,

f ′b2,v = 1 + w − wxy3 + 30(1 + w)2tx2z3,

fb3,v = 1 + w + 5xz3 − 6(1 + w)2wtx2y3.

Let us go through degree 10 in detail. The composition modulo (x, y, z)11 yields

x 7−→(

1 −700x4yz5

1 + w−

14x5y2z3

1 + w− 56(1 + w)2wx4y4z2

)

· x

y 7−→(

1 +175x4yz5

1 + w−

28x5y2z3

1 + w+ 56(1 + w)2wx4y4z2

)

· y

z 7−→(

1 +525x4yz5

1 + w+

42x5y2z3

1 + w

)

· z.


The occurring monomials are x4yz5 = tx3z4, x5y2z3 = t2x3z, x4y4z2 = t2x2y2, which have

directions (1, 4), (−2, 1) and (−1,−1), respectively. (The fact that the last one points in

the same direction as z explains why it does not show up in the last line: It belongs to

a wall acting trivially on z.) Thus we want to insert walls in directions (−1,−4), (2,−1)

and (1, 1) (10a, 10b and 10c in Figure 4.3) with functions

1 + atx3z4, 1 + bt2x3z, 1 + ct2x2y2,

for some a, b, c ∈ C[w]. We now explain how to determine the coefficients a, b, c. Since they

come in a product with monomials of degree 10, they do not interact with any monomial

of non-zero degree. Hence the influence to the composition can be computed from a

scattering diagram with only the wall in question and the three slabs bi with functions

fbi= 1 + w, i = 1, 2, 3. Here is the computation of the composition for the wall 10a:

xb17−→ x

b27−→ (1 + w)3x10a7−→

(

1 + atx3z4)4

· (1 + w)3x

b37−→(

1 + 4at(1 + w)−5x3z4)

· x.

Here we used(

1 + at(1 + w)−5x3z4)4

= 1 + 4at(1 + w)−5x3z4 modulo (x, y, z)10. The

second term has to cancel with −700(1 + w)−1tx3z4, which leads to a = 175(1 + w)4.

Said differently, we just have to commute the wall automorphism past all the auto-

morphisms of slabs in counterclockwise direction up to the reference cell. For the wall

labelled 10a there is only one such slab, b3, with normal vector (−3, 2), and the mono-

mial is x3z4 = z(1,4,0,4). The application of the associated automorphism yields the power

−⟨

(−3, 2), (1, 4)⟩

= −5 of 1 + w. The coefficient 4 = −⟨

(4,−1), (−1, 0)⟩

is picked up from

passing x past the wall in question. By the form of the automorphisms, looking at θkv (y)

or θkv (z) gives the same result. This is a consistency check.

The other coefficients can be computed analogously, giving

1 + 175(1 + w)4tx3z4, 1 + 14(1 + w)7t2x3z, 1 − 56(1 + w)4t2wx2y2

for the functions associated to the walls 10a, 10b and 10c, respectively.

Continuing in this fashion we can insert new walls essentially uniquely to make the

gluing consistent to any finite τ -order.

Remark 4.2. 1) There is one miracle here. Namely, while the scattering computation works

only after localization at 1+w, no negative powers of 1+w ever have to be inserted. This

is indispensable because walls and slabs can only carry regular functions. For general slab

functions and slab configurations this might indeed not be true, see the discussion in §4.4.1

of [GrSi2]. So we need a condition, and the condition we found to make it work ([GrSi2],


Definition 1.25,ii) morally says that the deformation problem has a unique solution locally

in codimension 2 up to isomorphism. This is a natural condition. Nevertheless, the proof

that no denominators occur in §4.4 of [GrSi2] is the most technical part of the story. We

refer to [GrSi2] for a more complete discussion.

2) For dim σj ∈ {n−1, n} the algorithm is the same, but there are higher degree monomials

with vanishing σj-order. In the (n − 1)-dimensional case a perturbation argument as in

§4.4 suffices to show that one can achieve consistency without changing the slab functions

and without introducing denominators, see [GrSi2] §4.3. In the n-dimensional case this

is not an issue because there are no slabs and hence there are no denominators and all

monomials have positive order anyway. This last case is essentially the situation that

was treated by Kontsevich and Soibelman in [KoSo]. In essence, our approach trades the

difficult gradient flow arguments in [KoSo] for algebraic arguments for the dimσj = n− 1

case, while the dim σj = n − 2 case is only non-trivial for n ≥ 3 and is the most difficult

issue.

4.4. The perturbation idea. Sometimes a simple geometric perturbation of the starting

data allows one to simplify a scattering situation. In [GrSi2], §4.2 we formalize this in the

notion of infinitesimal scattering diagrams. Here we just want to illustrate the concept by

one simple example.

Example 4.3. Consider the two-dimensional example of a local scattering situation with

xyz = t shown in Figure 4.4. By formally perturbing the slabs we arrive at Figure 4.5

x

y

z

v σ3

fρ1,v = 1 + x

fρ3,v = 1 + z

fρ2,v = 1 + yσ2

σ1

Figure 4.4. Scattering of three non-trivial slabs.

on the left. The functions are the slab functions on the various polyhedral pieces. While

this is not a legal polyhedral decomposition, we can still insert walls inductively to make

the diagram consistent to any finite order. The result is shown in Figure 4.5 on the right.

Note that this diagram is entirely determined by the most simple scattering situation of

§4.1 via affine isomorphisms.


1 + x

1 + yz

1 + x1 + x

1 + y

1 + y1 + z

1 + z

1 + z

1 + xy

1 + y

1 + xz

Figure 4.5. Perturbed scattering diagram.

Now the point of this discussion is that we can collapse this diagram again by super-

imposing walls and slabs with the same direction if necessary. For the present situation

1 + x

1 + y

1 + z

(1 + z)(1 + xy)

(1 + y)(1 + xz)

(1 + x)(1 + yz)

Figure 4.6. Result of collapsing Figure 4.5, right.

the result is shown in Figure 4.6. This configuration of slabs and walls is consistent to all

orders. �

We can also make this example projective. The smallest polarization produces a degen-

eration of cubic surfaces as follows.

Example 4.4. Consider the tropical manifold depicted in Figure 4.7, with three unit tri-

angles as maximal cells and one focus-focus singularity on each of the three interior edges.

The piecewise affine function is defined by the values 0, 0, 0, 1 on the 4 vertices U,X, Y,Z,

respectively. The vertices also define homogeneous coordinates written by the same sym-

bols. This exhibits X0 as the union of hyperplanes V (XY Z) ⊆ P3. Dehomogenizing with

respect to U gives the affine coordinates x0 = X/U , y0 = Y/U , z0 = Z/U , and in these

coordinates we define the slab functions by

fρ1,U = 1 + x0, fρ2,U = 1 + y0, fρ3,U = 1 + z0.


X

Y

Z

U

Figure 4.7. A tropical manifold for a degeneration of cubic surfaces.

Let us compute an affine chart for the deformation. To this end we look at Figure 4.6

and perform a computation similar to the basic gluing computation in §2.3. As reference

chamber choose the upper left quadrant in Figure 4.6. In this chamber x = X/U is the

uncorrected toric coordinate x0, but y = Y/U and z = Z/U receive corrections from

crossing walls and slabs. For z we need to cross the wall with function 1 + y0, while for y

there is the wall with function 1 + z0 and the slab with function (1 + x0)(1 + y0z0). The

complete list of relations therefore is

x0y0z0 = t

x = x0

z = (1 + y0)z0

y = (1 + x0 + z0 + y0z0 + x0y0z0)y0.

Eliminating x0, y0, z0 exhibits an affine chart for the deformation as the hypersurface

xyz = t(

(1 + t) + x + y + z)

.

in A4 with coordinates x, y, z, t. Homogenizing yields the degeneration of singular cubic

surfaces

XY Z = t(

(1 + t)U3 + (X + Y + Z)U2)

.

The general fibre has 3 ordinary double points at [1, 0, 0, 0], [0, 1, 0, 0] and [0, 0, 1, 0]; these

come from the vertices X,Y,Z on the boundary of B. �

5. Three-dimensional examples

So far we have essentially considered two-dimensional examples, the only exception be-

ing the three-dimensional sample scattering computation in §4.3. We already observed in

this example one complication in higher dimensions, the potential presence of poles in the

scattering procedure, see Remark 4.2,(1) for a discussion how this is handled. We also


observed in this example that the scattering procedure generally requires higher order

corrections to the slab functions, and the slab functions themselves propagate. In two

dimensions ad hoc solutions can be used to do this. In higher dimensions this propagation

is more complicated and we need a homological argument ([GrSi2], §3.5). Another, more

fundamental difference in higher dimensions is the fact that the codimension two intersec-

tion loci of walls and slabs, the joints, are higher dimensional. Thus in the automorphism

θj associated to a loop around a joint j there may be monomials tangent to j. These can

not be removed by inserting walls.

It nevertheless turns out that once we have run the homological argument, this only

happens for joints j contained in a codimension two cell σj of P, and the remaining

terms are either undirectional (pure t-powers, tl) or of the form tkzmρ

vv′/fρ,v ([GrSi2],

Proposition 3.23). Here mρvv′ is defined by affine monodromy as in (2.2). In other words,

these terms arise from pure t-powers on some other affine chart. The final step in our

algorithm uses a normalization procedure to get rid of these terms also ([GrSi2], §3.6).

The normalization essentially says that the logarithms of the slab functions do not contain

pure t-powers. Remarkably, this step not only makes sure that no obstructions arise in

the deformation process, but also makes t a canonical parameter in the sense of mirror

symmetry, traditionally denoted q.

Examples featuring all phenomena in this process are very complicated to run through

explicitly. We therefore content ourselves with one non-compact example treating a de-

generation of the total space of KP2, the so-called “local P2” from the mirror symmetry

literature [ChKlYa], and its mirror.

Example 5.1. Consider the tropical manifold B shown in Figure 5.1. There is a total of

six maximal cells σ0, . . . , σ5 with only σ0 bounded, a prism. The maximal cells adjacent

to the sides of the prism are σ1, σ2, σ3 as shown in the figure. The remaining σ4, σ5 are

adjacent to the bottom and top of σ0, respectively. The affine structure is then completely

determined by

σ0 = conv{

(0, 0, 0), (1, 0, 0), (0, 1, 0), (0, 0, 1), (1, 0, 1), (0, 1, 1)}

,

by the directions (2,−1,±1), (−1, 2,±1), (−1,−1,±1) of the unbounded edges in a chart

on the upper (plus sign) and lower (minus sign) half, and by requiring the monodromy

around the discriminant locus (the dashed lines in the figure) to be primitive and positive.

Note that (0, 0, 1) is an invariant tangent vector defining an affine projection as suggested

in Figure 5.1 on the right. As piecewise affine function ϕ we take a minimal one changing


X

U

W

σ0

σ2

σ1

σ3

V

Z

Y

Figure 5.1. The tropical manifold for a degeneration of KP2 (left) and its

projection to the plane (right).

slope by 1 along each codimension one cell and vanishing on σ0. In particular, on the six

unbounded edges ϕ takes the value 1 at the first integral point different from a vertex.

The nontrivial slabs (fb 6= 1) are the six vertical cells of codimension one. If we denote

by s the monomial with direction (0, 0,−1) and with σ0-order 0 and by s′ its inverse, the

gluing functions are

fb = 1 + s (upper half), fb = 1 + s′ (lower half).

Note that by (4.2) and the change of vertex formula (2.3) any one slab function determines

all the others.

This structure is already consistent along the three vertical edges of the prism. To get

consistency everywhere it only remains to insert six vertical walls in σ4 and σ5 each. These

extend the six vertical codimension one cells to infinity. There is no scattering because

the monomials carried by the walls and slabs all point in the same direction (0, 0, 1), the

invariant tangent vector, and hence the corresponding automorphisms mutually commute.

To study this example it is not advisable to write down the homogeneous coordinate

ring. In fact, because this is a non-compact example it does not suffice to take the six

vertices X,Y,Z,U, V,W as generators of a homogeneous coordinate ring. Rather we need

a number of generators of degree 0 defined by tangent vectors in the unbounded directions.

Due to the non-simplicial nature of the polyhedra this leads to a long list of generators

and an even longer list of relations.

Instead we use the construction via gluing affine patches from [GrSi2]. There are six

vertices, and correspondingly we have a cover of the degeneration by six affine open sets.


Figure 5.2 shows our choices of generators, where the arrows should be thought of as

r′

q′

u

z

y

w

p′r

q

u

z

y

w

p

Z

YX

W

v x v x

UV

Figure 5.2. The generators of the affine patches (left: bottom, right: top).

tangent to edges. Thus in a chart at U the generating monomials are

v = z(1,0,0,0), z = z(0,1,0,0), r = z(−1,−1,1,1), s = z(0,0,−1,0), t = z(0,0,0,1),

with single relation

rvzs = t,

a semi-stable (normal crossings) degeneration. Because of the symmetry of the example

the situation is analogous in the five other charts.

Now we have to adjust these local models by the slab functions. A computation analo-

gous to §2.3 gives

rvzs = (1 + s)t (at U), pwxs = (1 + s)t (at V ), quys = (1 + s)t (at W ),

r′vzs′ = (1 + s′)t (at X), p′wxs′ = (1 + s′)t (at Y ), q′uys′ = (1 + s′)t (at Z).

Here the variables take reference to any maximal cell containing them. Thus x, y, z, u, v, w

can all be thought of as monomials on σ0, while for example r is a monomial on σ2 or σ3.

The patching between these charts is given by affine geometry in the relevant maximal

cells. In particular, variables with the same name are all identified, and we have relations

such as

x = v−1, u = xw−1, s′ = s−1, q = pw3 etc.


As a consistency check one can verify compatibility of the relations with the gluing. For

example, going from the chart at U to the chart at V means the substitution

v 7→ x−1, z 7→ x−1w, r 7→ x3p, s 7→ s.

This maps the relation rvzs = (1 + s)t to pwxs = (1 + s)t, as expected. Similarly, to go

from U to X means substituting

v 7→ v, z 7→ z, r 7→ r′s′, s 7→ (s′)−1

into rvzs = (1 + s)t, leading to r′vzs′ = (1 + s′)t.

At this point we have written down a degeneration π : X → A1 = Spec C[t] with X

covered by six affine open sets. We claim that a general fibre Xt is an open subset of the

total space KP2 of the canonical bundle of P2. To this end fix t ∈ C \ {0} and define a

projection

κ : Xt −→ P2

by viewing the triples X,Y,Z or U, V,W as homogeneous coordinates on P2. Thus set-

theoretically the restriction of κ to the chart at U is

(r, v, z, s) 7−→ [1, v, z].

It is straightforward to check compatibility with the patching. For example, in the inter-

section with the chart at V we find

κ(p,w, x, s) = [x, 1, w] = [v−1, 1, v−1z] = [1, v, z].

Analogous computations show compatibility on the ring level.

The fibre of κ over a closed point of P2, say [1, v, z], is the hypersurface rvzs−(1+s)t = 0

in A2 = Spec(

C[r, s])

. Note that Xt is disjoint from s = 0 or from s′ = 0 since t 6= 0, so

it suffices to work in one chart only. If vz 6= 0 this is a hyperbola, hence isomorphic to

A1 \ {0}. On the other hand, if vz = 0 we must have s = −1 and r has no restrictions, so

this is an A1. The global meaning of this comes by observing that the fibre coordinates

p, q, r transform dually to the sections dx ∧ dw, dy ∧ du and dz ∧ dv of K2P. For example,

since y = z−1v and u = z−1,

qdy ∧ du = q(

− z−2vdz + z−1dv) ∧(

− z−2dz)

= qz−3dz ∧ dv = rdz ∧ dv.

Above we computed the fibre of κ over [1, v, z] to be given by (rvz−t)s = t. For given r, v, z

this has a solution s as long as rvz − t 6= 0. Thus rvz − t = 0 describes the hypersurface

locally that is being removed from KP2 to obtain Xt. Globally we are removing the graph

of a rational section of KP2 with poles along the toric divisor V (XYZ). �


Example 5.2. An example with somewhat complementary features to Example 5.1 is

provided by the mirror. Following the general recipe of [GrSi1] the tropical manifold to-

gether with the (multi-valued) function ϕ is obtained by a discrete version of the Legendre

transform ([GrSi1], Construction 1.15). By this construction the polyhedral decomposi-

tions of a tropical manifold and its mirror are combinatorially dual to each other. As this

is not primarily a paper about mirror symmetry we do not explain this construction, but

only state the result.

As shown in Figure 5.3 the tropical manifold is a double tetrahedron glued from six

Z

X

Y

W

X

Y

ZW

V

U

ρ2

ρ3

ρ1

Figure 5.3. The tropical manifold for the mirror of Example 5.1 (left)

and the central triangle containing the discriminant locus (right).

standard simplices. We have five exterior vertices labelled X,Y,Z,U, V and one interior

vertex W . In an affine chart at W the six emanating edges point in the directions of

the coordinate axes. The affine structure is then completely determined by requiring

the monodromy around the edges of the discriminant locus (dashed in the figure) to be

primitive and positive. The function ϕ is single-valued and can be taken to take value 0

on X,Y,U,W and value 1 on Z and V . This again has the property to change slope by 1

along each cell of codimension one.

There are three non-trivial slabs, the three horizontal triangles containing the discrim-

inant locus. Up to automorphisms there is only one set of gluing functions possible at

order 0, namely, in affine coordinates x = X/W , y = Y/W , z = Z/W at W :

fρ1,W = 1 + x + y, fρ2,W = 1 + y + z, fρ3,W = 1 + z + x.

The expressions at the other vertices follow from this by the change of vertex formula (2.3).


To make this structure consistent to all orders only requires propagating the slab func-

tions to the neighbouring slabs. This leads to

fb,W = 1 + x + y + z,

for any of the three slabs b = ρi.

The approach by homogeneous coordinates works well again in this case. We get the

toric relation

XY Z = tW 3,

and the homogenization of the gluing relation uv = (1 + x + y + z)t gives

UV = t2(X + Y + Z + W )W.

As one checks in local coordinates this is a complete set of relations. For t 6= 0 the

projective variety Xt is a conic bundle over Proj(

C[X,Y,Z,W ]/(XY Z − tW 3))

with

singular fibers over (X + Y + Z + W )W = 0. This base space of the conic bundle is the

quotient of P2 by the Z/3-action

ξ · [x0, x1, x2] = [ξx0, ξ2x1, x2],

for ξ a primitive third root of unity.

The suggestion in the literature for the mirror to KP2 is to take the family of non-

complete Calabi-Yau varieties defined by

uv = 1 + x + y + tx−1y−1

in C2 × (C∗)2 [ChKlYa]. Here u, v are the coordinates on C2. This is exactly the open

subset of our family fibering over the big cell of the weighted projective space. See [Gr1],

§4 for a discussion how this fits with the SYZ picture of local mirror symmetry.

As written this family does not come correctly parametrized for the purpose of mirror

symmetry. Rather, a period integral defines a new parameter q related to t by the so-

called mirror map. It is one striking feature of our approach that this mirror map comes

up naturally via the normalization condition. The present example is too local to illustrate

the need for doing this, but as mentioned at the beginning of this section, generally our

algorithm requires the logarithm of the slab functions to not contain any pure t-powers,

see [GrSi2], §3.6. In the present example this means adding a power series g =∑

l≥0 altl

to fb,W with the property that

log(

fb,W + g)

=∑

k≥1

(−1)k+1

k

(

x + y + z + g(xyz))k

∈ C[x, y, z℄


does not contain any monomials (xyz)l = tl. This condition determines the coefficients ak

inductively:

g(t) = −2t + 5t2 − 32t3 + 286t4 − 3038t5 + . . .

It follows from the period computations in [GbZa] that the modified family

XY Z = tW 3, UV = t2(

X + Y + Z + (1 + g(t))W)

W,

is then indeed written in canonical coordinates, that is, the mirror map becomes trivial. �

References

[ChKlYa] T.M. Chiang, A. Klemm, S.-T. Yau, E. Zaslow: Local mirror symmetry: calculations and inter-

pretations, Adv. Theor. Math. Phys. 3 (1999), 495–565.

[GbZa] T. Graber, E. Zaslow: Open-string Gromov-Witten invariants: calculations and a mirror “the-

orem”. in “Orbifolds in mathematics and physics (Madison, WI, 2001)”, 107–121, Contemp.

Math., 310, Amer. Math. Soc. 2002.

[Gr1] M. Gross: Examples of special Lagrangian fibrations, in: Symplectic geometry and mirror sym-

metry (Seoul, 2000), 81–109, World Sci. Publ. 2001

[Gr2] M. Gross: Toric Degenerations and Batyrev-Borisov Duality, Math. Ann. 333 (2005), 645–688.

[GrSi1] M. Gross, B. Siebert: Mirror symmetry via logarithmic degeneration data I, J. Differential

Geom. 72 (2006), 169–338.

[GrSi2] M. Gross, B. Siebert: From real affine to complex geometry, preprint arXiv:math/0703822,

128 pp.

[Ha] J. Harris: Algebraic geometry, Springer 1992.

[Ho] M. Hochster: Cohen-Macaulay rings, combinatorics and simplicial complexes, in: Ring theory II,

B.R. McDonald, R.A. Morris (eds.), Lecture Notes in Pure and Appl. Math. 26, M. Dekker 1977.

[KoSo] M. Kontsevich, Y. Soibelman: Affine structures and non-Archimedean analytic spaces, in: The

unity of mathematics (P. Etingof, V. Retakh, I.M. Singer, eds.), 321–385, Progr. Math. 244,

Birkhauser 2006.

[St] R. Stanley: Combinatorics and commutative algebra, Second ed., Birkhauser 1996.

[Sy] M. Symington: Four dimensions from two in symplectic topology, in: Topology and geometry of

manifolds (Athens, GA, 2001), 153–208, Proc. Sympos. Pure Math. 71, Amer. Math. Soc. 2003.

[Wi] J. Williamson: On the algebraic problem concerning the normal forms of linear dynamical sys-

tems, Amer. J. Math. 58 (1936), 141–163.

UCSD Mathematics, 9500 Gilman Drive, La Jolla, CA 92093-0112, USA

E-mail address: [email protected]

Department Mathematik, Universitat Hamburg, Bundesstraße 55, 20146 Hamburg, Ger-

many

E-mail address: [email protected]

an invitation to toric degenerations - arxiv · an invitation to toric degenerations mark gross and...

Documents